fix(ml): remove diversity hard gate, accept greedy mono-action policies

Greedy argmax legitimately converges to 1-2 actions when the model
is confident. The hard gate (unique_actions < 3 → 8.0 penalty) was
blocking ALL trials from using real backtest metrics, forcing FALLBACK.

Changes:
- Remove hard diversity short-circuit gate (unique_actions < 3)
- Reduce soft diversity penalty from 3.0 to 0.8 max (mild TPE signal)
- Fix early_stopping_enabled: false (was self.epochs > 12)
- Update test to validate mono-action policies produce good objectives

Multi-window backtest (3 windows, mean - 0.5*std) already handles
phantom Sharpe from single-bucket flukes.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

crates/ml/src/hyperopt/adapters/dqn.rs

@@ -2351,7 +2351,7 @@ impl HyperparameterOptimizable for DQNTrainer {
             // ALL trials to hit FALLBACK OBJECTIVE. With 8 epochs (~90s per trial),
             // it's cheaper to run all epochs and get real backtest metrics.
             // Long training runs (train-best, 50 epochs) still use early stopping.
-            early_stopping_enabled: self.epochs > 12,
+            early_stopping_enabled: false,
             q_value_floor: -5.0,
             min_loss_improvement_pct: 2.0,
             plateau_window: (self.epochs / 2).max(5),
@@ -3353,21 +3353,11 @@ impl HyperparameterOptimizable for DQNTrainer {
             // TRADE INSUFFICIENCY PENALTY: gives PSO gradient across degenerate plateau
             let trade_penalty = calculate_trade_insufficiency_penalty(backtest.total_trades);
 
-            // ACTION DIVERSITY PENALTY: scaled for 5-exposure DQN action space.
-            // - <3/5 unique exposures: degenerate trial, short-circuit with heavy penalty
-            //   (phantom metrics like Sharpe 2317, 100% WR from single exposure bucket)
-            // - 3-4/5: moderate penalty blended with composite score
-            // - 5/5: no penalty (healthy diversity — all exposures explored)
-            if backtest.unique_actions < 3 {
-                let diversity_ratio = backtest.unique_actions as f64 / 3.0;
-                // 8.0 for 1 action → ~2.67 for 2 actions (same scale as trade penalty)
-                let diversity_penalty = 8.0 * (1.0 - diversity_ratio);
-                info!(
-                    "DEGENERATE TRIAL (low diversity): unique_actions={}/5 → penalty={:.2} (ignoring Sharpe={:.2})",
-                    backtest.unique_actions, diversity_penalty + trade_penalty, backtest.sharpe_ratio
-                );
-                diversity_penalty + trade_penalty
-            } else if backtest.total_trades < 10 {
+            // ACTION DIVERSITY: greedy argmax legitimately converges to 1-2 actions
+            // when the model is confident in a directional bias. Multi-window backtest
+            // (3 independent windows, mean - 0.5*std) already penalizes lucky single-bucket
+            // flukes. Diversity is a soft signal, not a hard gate.
+            if backtest.total_trades < 10 {
                 info!(
                     "DEGENERATE TRIAL: {} trades → trade_penalty={:.2} (objective={:.2})",
                     backtest.total_trades, trade_penalty, trade_penalty
@@ -3388,12 +3378,11 @@ impl HyperparameterOptimizable for DQNTrainer {
                 // Component 2: HFT activity score (25% weight)
                 let hft_activity = calculate_hft_activity_score_wave10(buy_pct, sell_pct, hold_pct);
 
-                // Graduated diversity penalty for 3-4 unique exposures:
-                // 3 exposures → 3.0 penalty, 5 exposures → 0.0 penalty (linear ramp)
-                // 5 exposures: no penalty (all exposures explored)
+                // Mild diversity bonus: 5/5 unique → 0.0, 1/5 → 0.8 penalty.
+                // Soft signal for TPE, never dominates real backtest metrics.
                 let mid_diversity_penalty = if backtest.unique_actions < 5 {
-                    let ratio = (backtest.unique_actions as f64 - 3.0) / 2.0; // 0.0 at 3, 1.0 at 5
-                    3.0 * (1.0 - ratio.max(0.0))
+                    let ratio = backtest.unique_actions as f64 / 5.0;
+                    0.8 * (1.0 - ratio) // 0.64 for 1 action, 0.32 for 3, 0.0 for 5
                 } else {
                     0.0
                 };
@@ -4208,9 +4197,10 @@ mod tests {
     }
 
     #[test]
-    fn test_diversity_short_circuit_blocks_phantom_sharpe() {
-        // Verify that <10 unique actions produces a positive penalty objective,
-        // regardless of how high the Sharpe ratio is (phantom Sharpe=2317 bug)
+    fn test_low_diversity_adds_mild_penalty_not_hard_gate() {
+        // With greedy argmax, mono-action policies are valid. The diversity penalty
+        // should be a mild additive term (≤0.8), NOT a hard gate that overrides Sharpe.
+        // Multi-window backtest (3 windows, mean - 0.5*std) handles phantom Sharpe.
         let metrics = DQNMetrics {
             train_loss: 0.1,
             val_loss: 0.05,
@@ -4224,34 +4214,30 @@ mod tests {
             gradient_norm: 0.1,
             q_value_std: 0.1,
             backtest_metrics: Some(BacktestMetrics {
-                sharpe_ratio: 2317.0, // Phantom Sharpe!
-                win_rate: 100.0,
-                max_drawdown_pct: 0.0,
-                total_return_pct: 89000.0,
-                total_trades: 27000,
-                sortino_ratio: 22000.0,
-                calmar_ratio: 1_900_000.0,
-                var_95: 0.03,
-                cvar_95: 0.03,
-                beta: 0.0,
-                alpha: 0.0,
-                information_ratio: 0.0,
-                omega_ratio: 232000.0,
-                unique_actions: 2, // Only 2/5 — degenerate!
+                sharpe_ratio: 3.0, // Realistic good Sharpe
+                win_rate: 55.0,
+                max_drawdown_pct: 5.0,
+                total_return_pct: 12.0,
+                total_trades: 200,
+                sortino_ratio: 4.0,
+                calmar_ratio: 2.4,
+                var_95: -0.02,
+                cvar_95: -0.03,
+                beta: 0.1,
+                alpha: 0.05,
+                information_ratio: 1.5,
+                omega_ratio: 2.0,
+                unique_actions: 1, // Mono-action greedy — valid!
             }),
         };
 
         let objective = <DQNTrainer as HyperparameterOptimizable>::extract_objective(&metrics);
-        // Must be POSITIVE (bad) — the diversity short-circuit should override the composite
+        // With good Sharpe and 200 trades, objective should be NEGATIVE (good)
+        // even with mono-action. The 0.64 diversity penalty is far smaller than
+        // the composite score benefit from Sharpe=3.0.
         assert!(
-            objective > 0.0,
-            "Degenerate trial (2/5 exposures) should have positive objective, got {:.4}",
-            objective
-        );
-        // Should be in the 5-10 range (graduated penalty), not -369000
-        assert!(
-            objective < 20.0,
-            "Penalty should be moderate (5-10), got {:.4}",
+            objective < 0.0,
+            "Good Sharpe with mono-action should still produce negative (good) objective, got {:.4}",
             objective
         );
     }